"Agentic AI" describes a broader shift in how AI systems are designed: systems built to pursue a goal through a sequence of decisions and actions, rather than to answer a single prompt with a single response. It's a category and a design philosophy, distinct from "an AI agent," which is a specific system built within that philosophy. The distinction matters more than it sounds like it should.
What "agentic" actually describes
Agentic AI refers to AI systems characterized by agentic workflows: autonomous, multi-step reasoning and action, where the system plans, acts, observes the outcome, and adjusts, without a human specifying each individual step. This is a description of a behavioral pattern, not a specific product or piece of software.
What makes a workflow genuinely agentic
- Autonomy in sequencing: the system determines what steps are needed and in what order, rather than executing a human-predetermined sequence.
- Persistence toward a goal across multiple steps, not a single request-response exchange. A genuinely agentic system keeps working toward the outcome across however many steps that requires.
- Adaptation based on intermediate results, adjusting its plan when a step doesn't produce the expected outcome, rather than continuing blindly.
Why this is a spectrum, not a binary
Systems can be more or less agentic depending on how much autonomy they're given and how many steps they handle without human intervention. A system that plans three steps ahead and pauses for confirmation before each one is meaningfully agentic, but less so than one that completes an entire multi-step task independently. Thinking of "agentic" as a spectrum, rather than a label a system either has or doesn't, is a more accurate and useful way to reason about system design.
Why this term matters for how a project gets scoped
"We want agentic AI" is a design philosophy, not yet a specification. The real scoping questions are how much autonomy is appropriate for the specific task, what the failure modes look like if the system's autonomous reasoning goes wrong, and where a human-in-the-loop checkpoint is actually needed versus where it would just add friction without adding real safety. Answering those questions is what turns "we want agentic AI" into an actual buildable system.
How we approach this
We scope the degree of agentic autonomy a task actually needs, rather than defaulting to maximum autonomy because the term is prominent. We're explicit about where human checkpoints genuinely add safety and where they'd just slow the system down.